特斯拉加码机器人触觉，DeepSkin厚积十年，剑指薄膜式触觉新赛道..
Tesla Boosts Robot Tactility, DeepSkin Builds on Ten Years, Targeting the New Thin-Film Tactile Track..
Content Summary
The high-density electrodes behind a tactile thin film, self-developed tactile chip and manufacturing accumulation. Author | July Editor | Chestnut Oct 8, the US Patent and Trademark Office disclosed a Tesla humanoid robot patent application titled “THREE-DIM
High‑density electrodes behind a layer of tactile film, self‑developed tactile chip and manufacturing expertise.
The “Token economics” inside a ring.
Author｜July
Editor｜Chestnut
On October 8, the US Patent and Trademark Office disclosed a Tesla humanoid robot patent application titled “THREE-DIMENSIONAL SOFT COMPLIANT ARRAY TACTILE SENSOR WITH MULTI-MODAL SENSING AND SCALABLE MANUFACTURING METHODS” (Three‑dimensional soft compliant tactile sensor array with multi‑modal perception and scalable manufacturing methods). The front page of the patent includes an illustration—a robot hand with a sensor patch attached to the thumb, index finger, and palm,and the patches conform to the hand’s contour, as if a layer of skin were applied to the glove.
Illustration of the robot hand tactile patch (partial) from Tesla’s patent front page, source: US 2026/0310299 A1
But the detail worth paying more attention to is the latter half of the title“Scalable Manufacturing Method”.
A sensor patent puts “how to build” and “how to sense” into the same headline. When you look at Optimus’s situation over the past two years, it becomes clear. Musk has repeatedly said that a human‑level dexterous hand is the hardest part of Optimus, with difficulty “between a Model X and a Starship.” The Information also reported that Optimus’s hand faces challenges in manual assembly and the durability of tactile sensors.
The manufacturing issues involved in Tesla’s haptic patent share similarities with the problems that a number of Chinese flexible electronics companies have been repeatedly dealing with over the years.
DeepSkin is a flexible haptic brand under BianGe Technology and operated by Wuxi ShenJie Technology, and it is also one of the active players addressing this issue. Its parent company was founded in 2015 and has spent the past ten years building R&D and manufacturing capabilities for high‑density electrodes and flexible electronic products, holding more than 200 related patents and accumulating experience in automotive electronic product development and quality management, its high‑precision thin‑film electrode products have shipped over 1 billion yuan in total.
Nowadays, DeepSkin brings this accumulation into robotic tactile sensing, focusing on two directions: full‑palm tactile sensing for dexterous hands and large‑area electronic skin for robot bodies.On the product technology side, DeepSkin, leveraging high‑density thin‑film tactile arrays and its own tactile chip, has achieved tactile perception with high spatial resolution and high refresh rate.
Back toTesla’spatent, itshaptic solution andDeepSkin’sspecific devicesare not the same, but they both need to address some common questions:How to make the haptic interface conform to complex hand surfaces, capture subtle contacts, and incorporate both stability and manufacturability into the design.
1. Tesla haptic patent releases new signal,
Film-based sensing and mass manufacturing step into the spotlight
Tesla’s patent places flexible tactile perception and scalable manufacturing within the same technical solution for discussion.
Through flexible films, sensing arrays and the corresponding manufacturing processes, the sensor can adapt to contact areas of various shapes on a robot’s hand, converting external contact‑induced changes into tactile information usable by the robot.
In addition, the flexible sensing film must also conform to the shape of fingertips, finger pads, or palms, forming the corresponding curvature.The ultimate target for the film‑type sensor is a robotic hand that repeatedly grasps and contacts various objects.
This also means that film-type sensors must simultaneously address sensing and structural requirements: where forces are applied and how they are distributed require a sufficiently fine sensing array; conforming to curved surfaces and withstanding repeated contact require consideration of material and structural stability.
In these devices, high‑density electrodes provide the foundation for densely arranging sensing points. Metal mesh isone of the flexible electronics technologies on which DeepSkin is built, and the related experience can support the design of high‑density electrodes and flexible films. When it comes to palm‑scale tactile arrays, in addition to arranging a sufficient number of sensing points, it is necessary to consider the response consistency across different regions, as well as the impact of flexible materials, mounting structures, and long‑term use on sensing performance.
The engineering challenges faced by DeepSkin and Tesla share common ground: how to achieve high‑density sensing, maintain the stability of the thin‑film structure, conform to complex curvatures, and ensure that large‑scale sensor arrays deliver reliable, consistent responses.
Beyond these manufacturing experiences,the technology framework behind DeepSkin has also accumulated another tactile capability: reading minute signals from dense electrodes. Today, these in‑house chips are used to read thin‑film tactile arrays.
High-density electrodes, self-developedtactilechips andexisting flexible electronics industry ecosystem, constituting the starting point for DeepSkin’s entry into robotic tactile perception.
2. The next step for robotic dexterous hands,
is to give the entire palm a sense of touch
Why are robots now starting to take tactile sensing seriously?
Use a robot to pick up a paper cup as an illustration. Before the hand reaches out, the camera is responsible for determining where the cup is, its size, and its shape. When the fingers close, the problem changes: which positions the contact points land on, which side the force is applied to, and whether the cup slides down. At this point, most of the contact surface is blocked by the hand, and the camera cannot see it.
A six-axis force sensor on the wrist can tell the robot“how much total force is applied to the hand,” but the same resultant force could come from two fingers pinching or from the whole palm supporting it, resulting in completely different grasp postures. To distinguish these cases, a tactile array on the hand surface must provide a distribution map of “where the force is applied and how much.”
For past industrial robots, this wasn’t an issue—the workpiece’s position, orientation, and type are strictly constrained by fixtures, so the robot just follows the program; but embodied intelligence must operate in unstructured environments, where a cup on a table might be paper, glass, or even filled with water. Enumerating every possible contact scenario in advance is impossible; the robot must adjust on the fly during the motion. Tactile feedback fills in the post‑action information, reducing the robot’s need to guess the contact state.
In addition to providing immediate action feedback, haptics also have value at the data level. Video data can teach a model what a cup looks like, but it cannot teach how soft or slippery the cup is.The “2026 Embodied Intelligent Robot Sensor White Paper” jointly released by TÜV Rheinland and SENSOR CHINA points out that physical AI faces a shortage of high‑quality interactive data, while traditional video lacks interaction and feedback information. It can provide crucial data for learning robot manipulation strategies.
Current state of AI and sensor integration development, image source: “2026 Embodied Intelligent Robot Sensor White Paper”.
How much contact information can be recorded also depends on which parts the sensors cover.
Early industry approaches attached sensors to the fingertips and finger joints. However, real gripping rarely uses only the fingertips. Pinching the cup rim uses the fingertips, while gripping the cup body requires the finger pads and sides, and supporting from underneath relies mainly on the palm. As the motion progresses, the contact area shifts from the fingertips to the finger pads, then to the palm. Focusing only on the fingertips is equivalent to seeing only the beginning of this process.
SoDeepSkin primarily promotes “full‑palm haptics,” covering key contact areas such as fingertips, finger pads, finger joints, finger sides, and the palm surface.Leveraging a high‑density thin‑film haptic array,DeepSkin can design the sensing area, array specifications, and flexible structure according to different hand shapes and usage requirements.
In terms of fine perception, DeepSkin has achieved high‑resolution tactile arrays of 9,216 and 16,384 points, providing the technologyfoundation.
High-resolution texture recognition,Image source: Provided by DeepSkin
The array specifications differ across parts, which is an intentional design of DeepSkin: the fingertips need to discern texture and locate edges, requiring many spatial details; the palm mainly provides support, where coverage area is more important than density; the gripper’s contact area is a narrow strip, so the array is also shaped as a long strip. Allocating sensing resources according to contact tasks is far more efficient than uniformly covering the entire hand.
This means that for complete robot manufacturers, haptic suppliers need to be involved in structural design—Where the sensors are placed, which part of the hand each zone corresponds to, and how haptic data is temporally aligned with joint movements will all affect whether these data can be effectively used later. DeepSkin has formedcovering haptic sensors, self‑developed acquisition chips, acquisition hardware andfull‑stack technical capabilities of SDK/API, can provide robot OEMs with integrated hardware and software, supporting plug‑and‑play haptic solutions.
3. From high‑density electrodes to self‑developed haptic chip,
HapticsHow does the film achieve ultra‑high resolution?
Full‑foot coverage extends tactile sensation to more contact areas. The next challenge is to obtain sufficiently detailed contact information within limited thickness and space.
Vision‑tactile sensing is exemplified by GelSight. A tiny camera is placed at the fingertip to capture the deformation of a soft elastomer when it contacts an object, and algorithms are used to infer pressure, texture, and slip. It offers high resolution, and the resulting images are suitable for model training. The cost also comes from the optical system: the camera, illumination, and light path occupy space, the elastomer surface wears out, and each frame must be transmitted and processed. Extending these components to cover more of the hand requires further consideration of volume, layout, and maintenance.
DeepSkin captures details on a thin and soft contact surface using film-based tactile technology. The sensor film thickness for the dexterous hand and gripper direction can be controlled within 500 microns, supports a 2.5 mm bending radius, and the maximum size of a single sheet is about 2.2 m × 1.2 m. Different thicknesses and sizes correspond to different product configurations; specific values are subject to the respective model.
High‑density electrodes place more sensing points on the film, and the readout chip must capture the responses of these points. The flexible metal‑grid electrodes developed in‑housetouchsensingcapability, jointly supportingDeepSkin’s high‑resolution thin‑film tactile sensing.
The “entry fee” of high‑resolution thin‑film haptics, a considerable portion of which is paidduring the readout stage.
The array expands from 16×16 to 128×128, increasing the number of sensing points from 256 to 16,384—a 64‑fold rise. As the array size grows, the readout system must process more sensing information within a limited acquisition period, placing higher demands on the chip’s readout speed, channel architecture, and data‑processing capability.
In this regard,DeepSkin can leverage existing in‑house chip assets. According to the team,DeepSkin is built on proprietaryhapticchip and technology platform, supporting up to20,000 sensing points, with a chip‑level readout rate of up to 200 Hz, corresponding to a 5 ms acquisition cycle.The combination of high‑density electrodes and high‑speed readout provides core technical support for high‑resolution thin‑film haptics..Within the technology ecosystem built over the past decade, BianGe Technology has accumulated multiple mass‑produced, self‑developed chips, providing the technical foundation for DeepSkin to design products for various array sizes and acquisition requirements.
According to the R&D team,DeepSkin, based on a thin‑film tactile array and its own tactile chip, achieves high‑resolution arrays of 96×96 and 128×128,displayingfine tactile response imaging results.This demonstratesthe core of DeepSkincapabilities:On a thin and flexible contact surface,it captures more detailed tactile information.
High‑resolution texture recognition, imagesSource: provided by DeepSkin.
This accumulation also affects product development and cost. Toubao Research InstituteThe 2024 industry report estimated that chips account for30% of the electronic skin cost. Although this ratio cannot be directly applied to DeepSkin, it indicates that readout chips are also a significant cost in haptic products.
Upstream of the electronic skin industry chain. Image source: Toubao Research Institute “Overview of China’s Electronic Skin Industry: The Tactile Revolution of Robots”.
For DeepSkin, the value of developing its own tactile chip also lies in product design: the granularity of contact information needed, the update frequency, and how to integrate with hand structure must all be determined jointly by the array and the acquisition system. Developing these aspects together helps reduce repeated adaptations and provides a more concrete implementation path for high‑resolution thin‑film haptics.
4.Over 10 billion yuan in deliveries and automotive electronics experience,
How to support tactile mass production?
A piece of haptic skin is attached to the hand, and the impact of manufacturing quality is just beginning to show.
Imagine a scenario: the robot has already completed calibration based on the response of a sensor patch, and the grasping strategy has been tuned. Later, this skin wears out and needs to be replaced. With the same single press, the new skin’s readings are noticeably different, so the original calibration must be rechecked, and in severe cases the strategy also needs to be revalidated. If a customer deploys a hundred robots, this could happen a hundred times.
Therefore, for customers deploying at scale, consistency directly translates to operational costs. A recent in‑depth industry report put it bluntly: whether this chain can be considered successful depends on whether the same customer, using the same agile device, the second batch mass‑producedofwhether the data matches the first batch.
A massive magnetron sputtering coating system up to two stories tall, providing process support for the manufacturing of precise flexible electrodes for electronic skin. Image source: Provided by DeepSkin.
The flexible thin‑film electronics ecosystem that DeepSkin relies on has accumulated ten years of R&D, manufacturing, and delivery experience. According to the team,the company’s high‑precision thin‑film electrode products have shipped over 1 billion RMB in total,with customers spanning many well‑known domestic and international manufacturers.
In terms of manufacturing and quality management,with four production bases in Wuxi, Jiashan, Nanning, and Huizhou, about4.5ten thousand square meters of cleanroom, andwithIATF 16949 automotive industry quality management system certification as a representativethequality management system certification and experience in automotive product development and delivery.From precision manufacturing and production process control to product quality traceability, this manufacturing system, refined through long-term industrial practice, providesthe DeepSkin thin-film tactile sensor’sindustrial foundation for scalable manufacturing and mass delivery.
Image source: Provided by DeepSkin
For haptic products specifically, we also need to verify how each location responds under the specified load and whether the measured relationship changes after being mounted on a curved surface.Consistency concerns the differences between different devices, while stability concerns the changes of the same device over time.
For robot OEMs,manufacturing capability must ultimately be reflected in sustainable product delivery: Repeated production of the same design—can the response remain consistent across batches; if product deviations occur, can the cause be traced; after sensor wear replacement, can re‑tuning work be reduced.AndDeepSkinbehindthe quality management experience built from existing business,justprovides a foundation for addressing these issues.
In addition, soft materials also exhibit time‑dependent effects— When continuously compressed, deformation gradually increases, called creep; under a fixed deformation, stress gradually decreases, called stress relaxation; there is also hysteresis between loading and unloading. A robot picks up a part, waits, positions, and assembles, which may be a continuous action lasting several seconds. If the material changes subtly during this period, the robot may misjudge the contact state, e.g., thinking the gripper has loosened when actually the sensor has drifted. Whether the material has recovered by the start of the next grasp also affects the signal baseline.
Durability ultimately shows up in the customer’s cost. Replacing a skin involves downtime, disassembly, and recalibration. The Head Leopard Research Institute’s calculation model assumes that the value of electronic skin on a single humanoid robot will drop from about 9,200 yuan in 2024 to about 1,500 yuan in 2030. This data indicates that the industry’s expected market growth is based on the premise of “higher installation volume, lower unit price.”
DeepSkin’s modules are attached with backing adhesive, allowing whole-piece replacement, combined with factory calibration and batch traceability, aiming to make component replacement as simple as possible.
According to the technical lead of DeepSkin, its high‑resolution film‑type tactile sensor and its self‑developed tactile chip have formed a complete perception and acquisition system, and the related products have completed on‑site customer validation and are now entering a small‑batch production phase.
Image source: Provided by DeepSkin
From customer validation to mass production,DeepSkin’s capabilitiesextends farbeyond the sensor itself.High‑resolution perception, self‑developed tactile chips, system integration capabilities, and existing flexible electronics manufacturing experience together form the technical and industrialization foundation for its robot customers.
Robotic tactile sensing needs to continue improving resolution so that shapes, textures, and variations during contact are recorded more comprehensively; this information must also be fed into the control system in a timely manner and remain trustworthy after long‑term use and component replacement. It can be said that when Tesla wrote “scalable manufacturing” into the title of its tactile patent, it also brought manufacturing requirements beyond performance to the forefront.
From fingertips to the whole palm, from localized high‑resolution arrays to large‑area electronic skin,DeepSkin is building a thin‑film tactile product system for various robotic contact scenarios.Its differentiationfoundation lies not only inhigh‑density electrodes, proprietarytactilechips and full‑stack tactile technology capabilities,but also in the industrial support formed by more than RMB 1 billion in cumulative shipments of high‑precision thin‑film electrode products worldwide, four manufacturing bases, and an automotive‑grade quality management system.
(Cover image source: AI generated)
END.
Get MirrorSpeed